On this page

    Providers

    A provider is a record you add. The app ships no built-in provider and no default key. The protocol decides which role a record can fill:

    Protocol Role
    systemone, openai_decisions Decider
    openai_chat Generator
    openai_embeddings Embedder

    Four settings select the records:

    • triage.decider for triage.
    • judge.provider for the judge that proposes labels.
    • qa.provider for written answers.
    • embed.provider for semantic embeddings.

    A record used by a role cannot be deleted or changed to another kind until the role points elsewhere.

    Add a provider

    Open Settings, Providers and choose Add. The form asks for a label, a protocol, a base URL, a model and the fields that protocol needs. A decider also needs its question and token budgets. A generator needs both prices and an output cap. An embedder needs a complete embedding specification.

    The record id is a lower-case slug and does not change after creation. The provider file is providers.json in the data directory.

    Keys

    If you leave the secret name empty, the form suggests PROVIDER_<ID>_API_KEY. The value is written through the secrets route, never through the provider record. On Windows the secret is encrypted with DPAPI. On other platforms the secret store is read only, and the app reads MAILSENSE_SECRET_<NAME> from the environment.

    Test and model list

    Test makes one small real request: a one-question probe for a decider, a short chat call for a generator, or one short embedding. The result reports the stage that failed. A test is billable and is not covered by a budget.

    Fetch models sends GET {base_url}/models with the key you entered. The list is shown so you can apply a model; it is not stored. Missing limits and prices are left empty instead of invented.

    When no decider is configured

    Triage answers 422 provider_not_configured. Automatic triage stays idle and logs the reason. Search still works. The local classifier still learns from your answers, because it runs in the app and does not use a provider.

    A decider can be a remote service or a local OpenAI-compatible endpoint that you run yourself.

    What leaves the machine

    • Model discovery and the provider test, when you ask for them.
    • Triage decisions, including the sender, subject, a truncated body preview, recipients, folder, recent history and label examples.
    • Judge requests, which send message text to the configured generator.
    • Written answers, which send the retrieved messages the answer cites.
    • Embeddings, unless embed.provider points at the local runtime.

    Nothing else is sent. See privacy for the full request list.

    Source · AGPL-3.0-or-later · Documentation · Privacy